1-minute read of a 4 min report

Crafted ChatGPT Link Created a Scheduled Workspace Agent in Tests

By

1-minute read

What a defender needs to know

Zenity found that two Workspace Agents builder URL parameters could preselect a template and automatically submit attacker-written instructions in a logged-in user’s ChatGPT session. The victim needed Workspace Agents access and at least one previously authorized connector.

In Zenity’s proof of concept, one link drove the builder through connector attachment, approval changes, publishing, scheduling and a Preview run. The resulting agent checked the victim’s mailbox for attacker-supplied tasks and used connected services to return results. Zenity also demonstrated data searches, internal phishing and fraud lures in a controlled environment. It did not report exploitation against customers.

Zenity reported the flaw to OpenAI on June 4 and says it was fixed on June 8, before the research was published.

Workspace administrators should separate permission to use agents from permission to build or publish them. Review existing agents by owner, sharing state, connected app, authentication type, write-approval setting and schedule. Pay particular attention to recurring agents with write-capable email or collaboration connectors.

OpenAI publishes role-based controls for using, building and sharing agents, along with workspace and per-connector action limits. Use those controls to keep builder access narrow and connector reach specific. Include agents and shared connections in role-change and offboarding reviews.

Verify that every scheduled agent has a current owner, approved purpose and minimum connector permissions; disable any agent that cannot be accounted for.

Primary sources